Fig. 2

Berry detection and segmentation pipeline. A RGB image of a grapevine bunch acquired in the PhenoArch platform [8]. B Bounding boxes (red rectangles) detected by a Yolov4 deep-learning model trained to identify berries with at least 50% visible contour. C Vignettes cropped around the centre coordinates of detected boxes, and resized to 128 × 128 px. The resizing ensures that berries occupy a similar space in the vignette regardless of their size. D Binary segmentation masks predicted by a U-Net deep-learning model on berry vignettes. The model was trained to infer the shape of berries in the absence of occlusions. E Ellipse fitting of the contour points extracted from a segmentation mask, and projection of the ellipse (red) on the original image